Are there performance-portable programming frameworks for climate simulations across GPUs and CPUs?

Single-code performance portability is essential for the sustainable development of atmospheric models. In this talk, frameworks for single-code programming across CPUs and GPUs are evaluated using simple loops and a dynamical core of a cloud-resolving model. While OpenACC enables maintainable codes to perform reasonably on NVIDIA GPUs, it is not supported by AMD and Intel compilers. In contrast, the dynamical core in the OpenMP target suffered excessive overheads for atmospheric simulations on AMD and Intel GPUs. A simple loop in Kokkos ran several times slower than OpenMP on A64FX and Xeon CPUs. Currently, no framework appears to offer performance portability across processor vendors, making it unavoidable to focus on a limited set of targets.